Solving the Problem¶
Here i present my solution to this particular problem presented to me this afternoon¶
Lets check the importance of the variables¶
In [2]:
df_to_explore.info()
What about correlations?¶
In [3]:
sns.pairplot(data=df_to_explore)
Out[3]:
And maybe segmenting by Sales channel¶
In [4]:
sns.pairplot(data=df_to_explore,hue="Sales Channel")
Out[4]:
Gender?¶
In [5]:
sns.pairplot(data=df_to_explore,hue="Gender",palette="husl")
Out[5]:
The distribution of the Data looks all the same, therefore the is no difference at all between the different channels nor in the gender of the users.¶
Segmenting by state and sales channel¶
In [6]:
df_to_explore.groupby(["Sales Channel","State","Response"]).agg(Total=("Gender","count"))
Out[6]:
It is hard to see, let's plot it¶
In [7]:
df_to_explore.groupby(["State","Response"]).agg(Total=("Gender","count")).unstack(1)["Total"].plot(kind="bar")
plt.show()
We could also check the percentage of engagement segmenting by the offer¶
In [8]:
df_tmp=df_to_explore.groupby((["Renew Offer Type","Response"])).agg(Percentage=("Gender","count"))
df_tmp.groupby(level=0).apply(lambda x: 100*x / float(x.sum()))
Out[8]:
having more granularity we can check the type of car of the User¶
In [9]:
df_to_explore[df_to_explore["Response"]=="Yes"].groupby((["Vehicle Class","Renew Offer Type"])).agg(Total=("Gender","count"))["Total"].unstack().plot(kind="bar")
Out[9]:
In [10]:
df_to_explore[df_to_explore["Response"]=="Yes"].groupby((["Renew Offer Type", "Vehicle Class"])).agg(Total=("Gender","count"))["Total"].unstack().plot(kind="bar")
Out[10]:
We therefore conclude that the offer that performed the best was the 2 offer and the worts being the 4¶
Getting back to the Sales channel Performance¶
In [11]:
plt.figure(figsize=(9,7))
k=1
for i in df_to_explore["Sales Channel"].unique():
ax=plt.subplot(2,2,k)
df_tmp=df_to_explore[df_to_explore["Sales Channel"]==i].groupby(["Vehicle Size","Response"]).agg(Total=("Gender","count"))
df_tmp.groupby(level=0).apply(lambda x: 100*x / float(x.sum()))["Total"].unstack().plot(kind="bar",ax=ax, rot = 45, title ="% "+i)
plt.axhline(y=20,color="red",linestyle="--")
k+=1
plt.tight_layout()
plt.show()
There is a clear difference between the engagement present in different sales channel when consider the size of the vehicle. For all cases the "medsize" is the size that has better engagement for all channel and again we reinforce what we said before: The channel that performs the best is "Agent"¶
And What about a segmentation of our customers by their CLV and time since they started with our products?¶
In [13]:
plt.figure(figsize=(10,7))
k=1
for i in df_to_explore["CLV Range"].unique():
ax=plt.subplot(2,2,k)
df_to_explore[df_to_explore["CLV Range"]==i].groupby(["Months range","Response"]).agg(Total=("Gender","count"))["Total"].unstack().plot(kind="bar",ax=ax,title=i)
k+=1
plt.tight_layout()
plt.show()
For this case we see that people cataloged as "High_CLV" are the ones how performed the best. Something interesting to notice is that for the case of "Special CLV" there is a tendency from the new users as well as the Medium Users to Response "No" being this a sign of alarm to for the enterprise to take actions.¶
In order to complement our analysis i wanted to search for a model in which i could differentiate the costumers who answered yes and those who did not.¶
We will try to find through a hierarchical cluster the best description of our groups to then classify in a better way our Customers ( better insights about them)¶
Hirarchy¶
In [15]:
fig, ax = plt.subplots(1,1,figsize=(20, 12))
ax.set_title("Customer Dendograms")
dend = shc.dendrogram(shc.linkage(X, method='ward'),ax=ax)
ax.axhline(y=40,linewidth=3,linestyle='--',color='purple')
plt.show()
Who answered NO¶
In [17]:
x[x["Engaged"]==0].groupby("Cluster").mean()
Out[17]:
In [18]:
sns.pairplot(x[x["Engaged"]==0][['Customer Lifetime Value', 'Months Since Policy Inception',
'Income', 'Cluster']],hue="Cluster")
Out[18]:
Who answered Yes¶
In [19]:
x[x["Engaged"]==1].groupby("Cluster").mean()
Out[19]:
In [20]:
sns.pairplot(x[x["Engaged"]==1][['Customer Lifetime Value', 'Months Since Policy Inception',
'Income', 'Cluster']],hue="Cluster")
Out[20]:
We are able to fully characterise our clients who answered yes from those who did not.¶
- Moreover we are able to see the spectrum of the users who have answered affirmatively as well as the spectrum from those who did not.
But What is we want to make a difference between them, that is, what happen if we select the people who answered yes and then we try to understand its own segmentation.¶
Those who answered No¶
In [22]:
x.groupby("Cluster").mean()
Out[22]:
Let's visualize our results¶
In [23]:
sns.pairplot(x,hue="Cluster")
Out[23]:
Those who answered Yes¶
In [25]:
x.groupby("Cluster").mean()
Out[25]:
Let's visualize our results¶
In [26]:
sns.pairplot(x,hue="Cluster")
Out[26]:
